LFM2.5-24B-A2B is a 24B-parameter reasoning model with only 2B active parameters per token, designed for fast, private, on-device agents. It produces an explicit chain of thought before answering, calls tools, and works through multi-step tasks inside popular agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode. Because only 2B parameters activate per token, reasoning tokens are cheap enough to run agents continuously on local hardware. Both LFM2.5-24B-A2B and its base variant are open-weight on Hugging Face.

The architecture builds on LFM2-24B-A2B and keeps 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and GQA. The context window grows from 32K to 131K tokens, and an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens for better non-Latin tokenization. Unlike its predecessor, LFM2.5 is a reasoning-only model. Pre-training was scaled from roughly 17T to 33T tokens; mid-training then adds 2T tokens at 32K context weighted toward reasoning, math, code, instruction following, and multilingual text, plus 300B tokens at 131K context weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases. Post-training turns the base model into an agent in five stages: supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation, agentic distillation, and agentic reinforcement learning inside agent harnesses.

Liquid AI evaluated LFM2.5-24B-A2B against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math, and coding. It is the smallest model in the group, yet competitive with MoEs that have roughly 50% more total and active parameters; on LiveClaw it scores 48.09 versus 38.19 for Qwen3.6. It leads or matches the competition on five of the eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking, and leads on all three instruction-following benchmarks. Its AutomationBench score of 27.69 is more than three times the next-best model’s 8.72, and it posts 80.71 on ToolSandbox versus 75.02 for Qwen3.6. Larger models still hold an edge on knowledge-heavy benchmarks such as GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks such as SWE-Bench Verified and SciCode, so the source recommends pairing the model with retrieval or web search for knowledge-heavy workloads. All models were evaluated with vLLM using each benchmark’s recommended generation parameters.

On inference speed, the model ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX. It is the fastest of the compared MoEs on edge hardware, decoding 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, up to 1.7x faster than the next-best model while using the least memory. On a single NVIDIA B300 GPU it reaches over 21K output tokens per second at high concurrency. To use it, install transformers>=5.0.0, load the checkpoint from LiquidAI/LFM2.5-24B-A2B, and point an OpenAI-compatible backend at a local harness; default ports are 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang, and 1234 for LM Studio. The release also points to fine-tuning tutorials and agent harness configurations for OpenClaw and Pi.